Qwen4 Has Not Been Released Yet, but Code Already Reveals a 1M Context and an omni Model Name

Qwen4 hasn't been released yet, but the clues first appeared in Alibaba Qwen's own code. Mark Kretschmann pointed out on X that a merged update to the Qwen Code repository sets a 1M-token default context window for Qwen4 models, and the test fixture includes qwen4-omni-120b-think.

A context window can be understood as the maximum number of tokens a model can process at once. 1M tokens is a notable change for code repositories, long documents, and long conversations. The name qwen4-omni-120b-think looks more like preparation traces for a multimodal reasoning variant; omni usually means covering multiple input modalities.

The boundaries of the evidence also need to be made clear. Mark Kretschmann explicitly cautions that 1M is currently a client-side default, not a published model spec; this PR does not test an online Qwen4 endpoint; the code can show integration clues, but cannot show release timing.

One commenter, Dusan, believes the 1M default is more informative than the model name in the fixture. It is common to use made-up model names in tests, while a default context change is the real change. Cuth added more detailed clues: before the fix, every qwen4 id resolved to 262,144, and automatic compression would trigger at 22% of the real window; the same line sets qwen4-omni-120b-think to 1M, but the author says this id is the most uncertain; if its window is smaller, it is unclear whether the catalog or the exception row would provide the fallback. Aapakari reminded that the output limit has not changed; 1M input does not equal long output.

The multimodal part is also undecided. Someone asked whether the code implies native voice or strict vision, and Mark Kretschmann replied that the code does not say.

People have also done the math for local deployment. Own Your Compute estimates that if qwen4-omni-120b-think really exists, Q4 quantization would be about 65 to 70GB, not including KV cache. Mac Studio and Spark users can handle it; GPU machines would need three to four 24GB cards. With a 1M default context, KV cache will become the main battleground for VRAM, while weights fall into the background.

What can be confirmed now is a concrete integration clue, still some distance from an official release spec. The test fixture leaked the model name first; code runs ahead of blog posts, and Qwen's open-source cadence has once again left the suspense to the commit history.

Source: Mark Kretschmann's original tweet